diff --git a/litellm/integrations/langfuse/langfuse_otel.py b/litellm/integrations/langfuse/langfuse_otel.py index 34d2c9fce0d..ee75e6501e8 100644 --- a/litellm/integrations/langfuse/langfuse_otel.py +++ b/litellm/integrations/langfuse/langfuse_otel.py @@ -35,6 +35,8 @@ from litellm.types.utils import StandardCallbackDynamicParams if TYPE_CHECKING: from opentelemetry.trace import Span as _Span + from litellm.proxy._types import UserAPIKeyAuth + Span = _Span else: Span = Any @@ -126,6 +128,26 @@ def _starts_its_own_trace(span: object, kwargs: Mapping[str, object]) -> bool: ) +def _a_destination_records_the_call(kwargs: Mapping[str, object]) -> bool: + """Whether a key's or team's ``langfuse_otel`` destination gets this call from the OTel v2 fan-out, which records + every call but one whose logging payload shows it never reached a provider.""" + from litellm.integrations.otel.model.metadata import LLMCallEvent + + if not _a_destination_carries_the_trace(): + return False + call: Final = LLMCallEvent.from_dict(kwargs) + return call.payload is None or call.reached_a_provider + + +def _a_destination_carries_the_trace() -> bool: + """Whether this request resolved a key's or team's ``langfuse_otel`` destination, whose fan-out sends it the whole + trace.""" + from litellm.integrations.otel.model.config import ExporterOwner + from litellm.integrations.otel.plumbing.context import destination_backends + + return ExporterOwner.LANGFUSE_OTEL in destination_backends() + + class LangfuseOtelLogger(OpenTelemetry): def __init__(self, config=None, *args, **kwargs): # Prevent LangfuseOtelLogger from modifying global environment variables by constructing config manually @@ -541,6 +563,27 @@ class LangfuseOtelLogger(OpenTelemetry): """ return None + def _emit_once( + self, + kwargs: dict[str, object], # mutable-ok: the v1 dedupe marker is written into the request's kwargs + *scope: object, + ) -> bool: + """Whether to emit this span: once per request, and never for a call a Langfuse destination records.""" + return not _a_destination_records_the_call(kwargs) and super()._emit_once(kwargs, *scope) # pyright: ignore[reportUnknownMemberType] # the v1 hook takes a bare dict + + async def async_post_call_failure_hook( + self, + request_data: dict[str, object], # mutable-ok: the v1 hook contract + original_exception: Exception, + user_api_key_dict: "UserAPIKeyAuth", + traceback_str: str | None = None, + ) -> None: + """Record the proxy failure of a request no Langfuse destination carries.""" + if not _a_destination_carries_the_trace(): + await super().async_post_call_failure_hook( # pyright: ignore[reportUnknownMemberType] # the v1 hook takes a bare dict + request_data, original_exception, user_api_key_dict, traceback_str + ) + async def async_service_success_hook(self, *args, **kwargs): """ Langfuse should not receive service success logs. diff --git a/litellm/integrations/otel/emitter.py b/litellm/integrations/otel/emitter.py index c02ea6032a0..87a432ee7ce 100644 --- a/litellm/integrations/otel/emitter.py +++ b/litellm/integrations/otel/emitter.py @@ -13,9 +13,9 @@ from opentelemetry.trace.status import Status, StatusCode from litellm.integrations.otel.mappers import resolve_mappers from litellm.integrations.otel.mappers.base import AttributeMapper, AttrValue, SpanData -from litellm.integrations.otel.mappers.langfuse import LANGFUSE_REQUEST_METADATA_ATTRIBUTES +from litellm.integrations.otel.mappers.langfuse import LANGFUSE_REQUEST_METADATA_ATTRIBUTES, LangfuseMapper from litellm.integrations.otel.mappers.openinference import fit_indexed_messages -from litellm.integrations.otel.model.config import OpenTelemetryV2Config +from litellm.integrations.otel.model.config import ExporterOwner, OpenTelemetryV2Config from litellm.integrations.otel.model.payloads import ( GuardrailSpanData, LLMCallSpanData, @@ -34,6 +34,7 @@ from litellm.integrations.otel.model.spans import ( mcp_tool_call_span_name, service_span_name, ) +from litellm.integrations.otel.plumbing.context import destination_backends from litellm.integrations.otel.plumbing.events import GenAIEventRecorder from litellm.integrations.otel.plumbing.providers import to_otel_span_kind @@ -168,6 +169,9 @@ class SpanEmitter: self._mappers: list[AttributeMapper] = ( list(mappers) if mappers is not None else resolve_mappers(config.mapper_names) ) + self._langfuse: Final = ( + None if any(isinstance(mapper, LangfuseMapper) for mapper in self._mappers) else LangfuseMapper() + ) # Bounded LRU (ordered by insertion / most-recent touch). Storing keys # only — the value is unused — so it behaves like a capped set. self._emitted: OrderedDict[tuple[str, SpanRole], None] = OrderedDict() @@ -295,17 +299,27 @@ class SpanEmitter: ) else None ) - mapped: Final = MappingProxyType( - {key: value for mapper in self._mappers for key, value in mapper.map(data).items()} + configured: Final = {key: value for mapper in self._mappers for key, value in mapper.map(data).items()} + for_langfuse: Final = ( + dict( + sorted( + self._langfuse.map(data).items(), + key=lambda item: item[0] not in LANGFUSE_REQUEST_METADATA_ATTRIBUTES, + ) + ) + if self._langfuse is not None and ExporterOwner.LANGFUSE_OTEL in destination_backends() + else _NO_ATTRIBUTES ) + mapped: Final = MappingProxyType({**configured, **for_langfuse}) stamped_later: Final = error_attributes(error) if error else _NO_ATTRIBUTES reserved: Final = len(stamped_later.keys() - mapped.keys()) overwrites: Final = carried_keys(span) & mapped.keys() - # The Langfuse request metadata sheds first, so a span at its attribute limit loses nothing else to it. + # The Langfuse request metadata, and the Langfuse vocabulary only a destination asked for, shed first, + # so a span at its attribute limit loses nothing else to them. fitted: Final = fit_indexed_messages( MappingProxyType({key: value for key, value in mapped.items() if key not in overwrites}), attribute_budget(span, reserved), - LANGFUSE_REQUEST_METADATA_ATTRIBUTES, + LANGFUSE_REQUEST_METADATA_ATTRIBUTES.union(for_langfuse), ) for key, value in mapped.items(): if key in overwrites or key in fitted: diff --git a/litellm/integrations/otel/logger.py b/litellm/integrations/otel/logger.py index 4c5e65ff07d..4c9618800ab 100644 --- a/litellm/integrations/otel/logger.py +++ b/litellm/integrations/otel/logger.py @@ -541,9 +541,7 @@ class OpenTelemetryV2(CustomLogger): # gate rejection carries ``is_no_upstream_call`` and gets no span. A cache # hit never hands off to a provider, so ``cache_hit`` substitutes for the # upstream stamp, since the payload proves the call completed. - if carrier is None and ( - call.is_no_upstream_call or call.payload is None or (not call.upstream_started and not call.cache_hit) - ): + if carrier is None and not call.reached_a_provider: return None try: return self._finish_carrier(carrier, call, start_time, end_time) diff --git a/litellm/integrations/otel/mappers/langfuse.py b/litellm/integrations/otel/mappers/langfuse.py index b55d9746f63..5aed173254c 100644 --- a/litellm/integrations/otel/mappers/langfuse.py +++ b/litellm/integrations/otel/mappers/langfuse.py @@ -103,6 +103,18 @@ class LangfuseMapper: ), } + #: Every attribute this mapper can write on a model call. + LLM_CALL_ATTRIBUTES: Final[frozenset[str]] = frozenset( + { + *_LLM_CALL_ATTRS, + *_BLOB_ATTRS, + LANGFUSE_TRACE_NAME, + LANGFUSE_TRACE_USER_ID, + LANGFUSE_TRACE_SESSION_ID, + LANGFUSE_TRACE_TAGS, + } + ) + def map(self, data: SpanData) -> AttributeMap: match data: case LLMCallSpanData(): diff --git a/litellm/integrations/otel/model/metadata.py b/litellm/integrations/otel/model/metadata.py index a9a57de04ac..28d6ad54607 100644 --- a/litellm/integrations/otel/model/metadata.py +++ b/litellm/integrations/otel/model/metadata.py @@ -272,6 +272,11 @@ class LLMCallEvent: trace: TraceControls session_id: str | None + @property + def reached_a_provider(self) -> bool: + """Whether the call was handed to a provider or answered from the response cache.""" + return not self.is_no_upstream_call and self.payload is not None and (self.upstream_started or self.cache_hit) + @classmethod def from_dict(cls, kwargs: Mapping[str, object]) -> LLMCallEvent: raw_payload: Final = kwargs.get("standard_logging_object") diff --git a/litellm/integrations/otel/plumbing/providers.py b/litellm/integrations/otel/plumbing/providers.py index a369e352609..39e78194228 100644 --- a/litellm/integrations/otel/plumbing/providers.py +++ b/litellm/integrations/otel/plumbing/providers.py @@ -47,6 +47,7 @@ from litellm.integrations.otel.mappers.langfuse import ( LANGFUSE_OBSERVATION_OUTPUT, LANGFUSE_REQUEST_METADATA_ATTRIBUTES, LANGFUSE_TRACE_NAME, + LangfuseMapper, ) from litellm.integrations.otel.model.config import ( CaptureMessageContent, @@ -490,19 +491,25 @@ def _is_litellm_span(span: ReadableSpan) -> bool: return scope is not None and scope.name == LITELLM_TRACER_NAME -def without_langfuse_request_metadata(span: ReadableSpan) -> ReadableSpan: - """``span`` without the request metadata the Langfuse mapper writes for a Langfuse destination. +def without_destination_langfuse_attributes(span: ReadableSpan, langfuse_mapped: bool) -> ReadableSpan: + """``span`` without the Langfuse attributes only a Langfuse destination asked for. - Only LiteLLM's own spans lose it: any other span keeps every key its own code set. + That is the request metadata the Langfuse mapper writes and, on a model call written by + a logger whose own mappers leave the Langfuse vocabulary out (``langfuse_mapped`` off), + every Langfuse attribute, since that logger wrote them for a key's or team's Langfuse + destination alone. Only LiteLLM's own spans lose them: any other span keeps every key + its own code set. """ attributes: Final = span.attributes or _NO_ATTRIBUTES - if not (_is_litellm_span(span) and any(key in LANGFUSE_REQUEST_METADATA_ATTRIBUTES for key in attributes)): + dropped: Final = ( + LANGFUSE_REQUEST_METADATA_ATTRIBUTES + if langfuse_mapped or not is_llm_call_span(span) + else LangfuseMapper.LLM_CALL_ATTRIBUTES + ) + if not (_is_litellm_span(span) and any(key in dropped for key in attributes)): return span return _CountedSpanView( - span, - MappingProxyType( - {key: value for key, value in attributes.items() if key not in LANGFUSE_REQUEST_METADATA_ATTRIBUTES} - ), + span, MappingProxyType({key: value for key, value in attributes.items() if key not in dropped}) ) @@ -658,6 +665,7 @@ def _for_destination( span: ReadableSpan, destination: "OtelDestination", default_capture: str, + langfuse_mapped: bool, ) -> ReadableSpan: """The view of ``span`` a tenant destination receives. @@ -672,13 +680,16 @@ def _for_destination( and the query string of the request URL, which can hold the same key. A destination whose effective setting disables span content loses the prompt and response content too, and a destination for any backend but Langfuse loses the - request metadata the Langfuse mapper writes on LiteLLM's own spans. The span itself - stays, so the tenant still gets the whole trace tree. + Langfuse attributes only a Langfuse destination asked for (see + :func:`without_destination_langfuse_attributes`). The span itself stays, so the + tenant still gets the whole trace tree. """ extra: Final = destination.resource_attributes defaults: Final = destination.resource_defaults source: Final = ( - span if destination.callback_name == ExporterOwner.LANGFUSE_OTEL else without_langfuse_request_metadata(span) + span + if destination.callback_name == ExporterOwner.LANGFUSE_OTEL + else without_destination_langfuse_attributes(span, langfuse_mapped) ) attributes: Final = source.attributes or _NO_ATTRIBUTES database: Final = _is_database_span(attributes) @@ -728,8 +739,10 @@ class TenantFanOutSpanProcessor(SpanProcessor): default_capture_by_backend: Mapping[str, str] = MappingProxyType({}), pending_drains: int = _MAX_PENDING_DRAINS, drain_pool: _DrainPool | None = None, + langfuse_mapped: bool = True, ) -> None: self._operator_sinks: Final = operator_sinks + self._langfuse_mapped: Final = langfuse_mapped self._excluded_db_systems: Final = excluded_db_systems self._default_capture: Final = default_capture self._default_capture_by_backend: Final = default_capture_by_backend @@ -822,7 +835,8 @@ class TenantFanOutSpanProcessor(SpanProcessor): try: processor.on_end( _scoped( - _for_destination(span, destination, self._default_capture_for(destination)), destination.span_scope + _for_destination(span, destination, self._default_capture_for(destination), self._langfuse_mapped), + destination.span_scope, ) ) except Exception as exc: # noqa: BLE001 # one destination's failure must not cost the others their span @@ -1198,17 +1212,19 @@ class _OverriddenBackendFilter(SpanProcessor): return self._inner.force_flush(timeout_millis) -class _LangfuseRequestMetadataFilter(SpanProcessor): - """Strip the Langfuse request metadata from LiteLLM's own spans before ``inner`` exports them.""" +class _DestinationLangfuseAttributeFilter(SpanProcessor): + """Strip the Langfuse attributes only a Langfuse destination asked for from LiteLLM's own spans before ``inner`` + exports them.""" - def __init__(self, inner: SpanProcessor) -> None: + def __init__(self, inner: SpanProcessor, langfuse_mapped: bool) -> None: self._inner: Final = inner + self._langfuse_mapped: Final = langfuse_mapped def on_start(self, span: SDKSpan, parent_context: Context | None = None) -> None: self._inner.on_start(span, parent_context) def on_end(self, span: ReadableSpan) -> None: - self._inner.on_end(without_langfuse_request_metadata(span)) + self._inner.on_end(without_destination_langfuse_attributes(span, self._langfuse_mapped)) def shutdown(self) -> None: self._inner.shutdown() @@ -1491,13 +1507,18 @@ def _spec_processor( spec: ExporterSpec, use_simple_processor: bool | None, strip_message_content: bool, - strip_langfuse_request_metadata: bool, + strip_destination_langfuse_attributes: bool, + langfuse_mapped: bool, ) -> SpanProcessor: exporting: Final = _processor_for( _exporter_from_spec(spec), (spec.use_simple_processor if spec.use_simple_processor is not None else use_simple_processor), ) - scoped: Final = _LangfuseRequestMetadataFilter(exporting) if strip_langfuse_request_metadata else exporting + scoped: Final = ( + _DestinationLangfuseAttributeFilter(exporting, langfuse_mapped) + if strip_destination_langfuse_attributes + else exporting + ) return _MessageContentFilter(scoped) if strip_message_content else scoped @@ -1548,8 +1569,9 @@ def build_tracer_provider( ``config.langfuse_span_scope`` narrows the exporter owned by ``langfuse_otel`` alone; a collector or any other backend in the same config keeps the full tree. - In a config that writes the Langfuse vocabulary, every exporter but ``langfuse_otel``'s - goes without the request metadata the Langfuse mapper writes on LiteLLM's own spans. + Every exporter but ``langfuse_otel``'s goes without the Langfuse attributes only a + Langfuse destination asked for (see :func:`without_destination_langfuse_attributes`), + which any logger writes for a request with a Langfuse destination. ``strip_message_content`` removes prompt and response content before every exporter, for a routed provider whose spans can carry content another destination asked for. @@ -1564,11 +1586,9 @@ def build_tracer_provider( provider.add_span_processor(_processor_for(exporter, use_simple_processor)) return provider + langfuse_mapped: Final = _maps_langfuse(config) # ``config._normalize`` guarantees at least one spec (it folds the top-level # ``exporter``/``endpoint``/``headers`` fields in when ``exporters`` is empty). - writes_langfuse: Final = "langfuse" in config.mapper_names or any( - spec.owner is ExporterOwner.LANGFUSE_OTEL for spec in config.exporters - ) for spec in config.exporters: if spec.requires_headers and not spec.headers: continue @@ -1576,7 +1596,8 @@ def build_tracer_provider( spec, use_simple_processor, strip_message_content, - strip_langfuse_request_metadata=writes_langfuse and spec.owner is not ExporterOwner.LANGFUSE_OTEL, + strip_destination_langfuse_attributes=spec.owner is not ExporterOwner.LANGFUSE_OTEL, + langfuse_mapped=langfuse_mapped, ) owner = spec.owner.value if tenant_overrides and spec.owner is not None else None scope = _operator_scope(config, spec) @@ -1613,8 +1634,9 @@ def attach_tenant_fan_out( so exactly one fan-out lands. ``configs`` name the operator's own exporters, one config per v2 logger since each keeps its own provider and still writes its account, so an additive destination pointing at any of them is delivered once - rather than twice. ``excluded_db_systems`` only filters what the fan-out - delivers, never the operator's own exporters. + rather than twice. The first is the config of the logger that owns ``provider``, + whose model calls the fan-out carries. ``excluded_db_systems`` only filters what + the fan-out delivers, never the operator's own exporters. """ with _FAN_OUT_ATTACH_LOCK: if any(isinstance(processor, TenantFanOutSpanProcessor) for processor in _attached_processors(provider)): @@ -1625,10 +1647,16 @@ def attach_tenant_fan_out( excluded_db_systems=excluded_db_systems, default_capture=_operator_capture_mode(configs), default_capture_by_backend=MappingProxyType(dict(_owned_capture_modes(configs))), + langfuse_mapped=any(_maps_langfuse(config) for config in configs[:1]), ) ) +def _maps_langfuse(config: OpenTelemetryV2Config) -> bool: + """Whether a logger built from ``config`` writes the Langfuse vocabulary on every model call.""" + return "langfuse" in config.mapper_names + + def _operator_capture_mode(configs: Sequence[OpenTelemetryV2Config]) -> str: """The mode of the operator's generic ``otel`` callback, so ``callback_settings.otel`` counts as the global default""" return next( diff --git a/litellm/integrations/otel/presets/destinations.py b/litellm/integrations/otel/presets/destinations.py index a705fc16bbc..673337854ce 100644 --- a/litellm/integrations/otel/presets/destinations.py +++ b/litellm/integrations/otel/presets/destinations.py @@ -15,7 +15,7 @@ import litellm from litellm._logging import verbose_logger from litellm.integrations.otel.model.destination import OtelDestination from litellm.litellm_core_utils.url_utils import is_url_destination_allowed_by_host -from litellm.types.utils import OtelSpanScope, StandardCallbackDynamicParams +from litellm.types.utils import CaptureMessageContent, OtelSpanScope, StandardCallbackDynamicParams #: An endpoint plus the OTLP transport to reach it with, or ``None`` when the backend #: names no destination. The transport is ``None`` where the backend has only one. @@ -111,7 +111,7 @@ _REQUIRED_HEADERS_BY_CALLBACK: Final[Mapping[str, frozenset[str]]] = MappingProx _NO_ATTRS: Final[Mapping[str, str]] = MappingProxyType({}) -def _arize_project() -> Mapping[str, str]: +def _arize_project(_params: StandardCallbackDynamicParams) -> Mapping[str, str]: from litellm.integrations.arize.arize import ArizeLogger project: Final = ArizeLogger.get_arize_config().project_name @@ -124,10 +124,27 @@ def _arize_default_project() -> Mapping[str, str]: return MappingProxyType({"model_id": ARIZE_DEFAULT_PROJECT}) -#: Resource attributes the operator chose for a backend, applied to every export the -#: team receives over the span's own resource: ``ARIZE_PROJECT_NAME`` for Arize. -_RESOURCE_ATTRIBUTES_BY_CALLBACK: Final[Mapping[str, Callable[[], Mapping[str, str]]]] = MappingProxyType( - {"arize": _arize_project} +def _stands_in_for_the_legacy_langfuse_logger() -> bool: + """Whether a Langfuse destination replaces the older Langfuse logger's copy, which a proxy without Langfuse keys + of its own sends for a key's or team's ``langfuse_otel``.""" + return not (os.environ.get("LANGFUSE_PUBLIC_KEY") and os.environ.get("LANGFUSE_SECRET_KEY")) + + +def _langfuse_environment(params: StandardCallbackDynamicParams) -> Mapping[str, str]: + """The key's or team's ``langfuse_environment``, else ``LANGFUSE_TRACING_ENVIRONMENT``, where the older Langfuse + logger wrote it.""" + from litellm.types.integrations.langfuse_otel import LangfuseSpanAttributes + + if not _stands_in_for_the_legacy_langfuse_logger(): + return _NO_ATTRS + environment: Final = params.get("langfuse_environment") or os.environ.get("LANGFUSE_TRACING_ENVIRONMENT") + return MappingProxyType({LangfuseSpanAttributes.LANGFUSE_ENVIRONMENT.value: environment} if environment else {}) + + +#: Resource attributes a backend gets on every export the team receives, over the +#: span's own resource: ``ARIZE_PROJECT_NAME`` for Arize, the Langfuse environment. +_RESOURCE_ATTRIBUTES_BY_CALLBACK: Final[Mapping[str, Callable[[StandardCallbackDynamicParams], Mapping[str, str]]]] = ( + MappingProxyType({"arize": _arize_project, "langfuse_otel": _langfuse_environment}) ) #: Resource attributes a backend needs on every export to accept it, filled only where @@ -137,12 +154,14 @@ _RESOURCE_DEFAULTS_BY_CALLBACK: Final[Mapping[str, Callable[[], Mapping[str, str ) -def _resource_attributes(callback_name: str, service_name: str | None) -> Mapping[str, str]: +def _resource_attributes( + callback_name: str, service_name: str | None, params: StandardCallbackDynamicParams +) -> Mapping[str, str]: backend: Final = _RESOURCE_ATTRIBUTES_BY_CALLBACK.get(callback_name) return MappingProxyType( { **({"service.name": service_name} if service_name else {}), - **(backend() if backend is not None else {}), + **(backend(params) if backend is not None else {}), } ) @@ -152,6 +171,14 @@ def _resource_defaults(callback_name: str) -> Mapping[str, str]: return backend() if backend is not None else _NO_ATTRS +def _capture_message_content(callback_name: str, configured: str | None) -> str | None: + """The key's or team's own setting, else the prompt and response where the older Langfuse logger sent them + whatever the global setting.""" + if configured is not None or callback_name != "langfuse_otel" or not _stands_in_for_the_legacy_langfuse_logger(): + return configured + return CaptureMessageContent.SPAN_ONLY + + def _span_scope(callback_name: str, params: StandardCallbackDynamicParams) -> OtelSpanScope: if callback_name != "langfuse_otel": return "full" @@ -214,12 +241,12 @@ def destination_for( return OtelDestination( endpoint=endpoint, headers=MappingProxyType(dict(headers)), - resource_attributes=_resource_attributes(callback_name, service_name), + resource_attributes=_resource_attributes(callback_name, service_name, params), resource_defaults=_resource_defaults(callback_name), callback_name=callback_name, protocol=protocol, span_scope=_span_scope(callback_name, params), - capture_message_content=capture_message_content, + capture_message_content=_capture_message_content(callback_name, capture_message_content), excluded_db_systems=excluded_db_systems, success_sampling_rate=success_sampling_rate, error_sampling_rate=error_sampling_rate, diff --git a/tests/integration/observability/test_langfuse_otel_legacy_destinations.py b/tests/integration/observability/test_langfuse_otel_legacy_destinations.py new file mode 100644 index 00000000000..aaff856fda8 --- /dev/null +++ b/tests/integration/observability/test_langfuse_otel_legacy_destinations.py @@ -0,0 +1,224 @@ +"""Without operator Langfuse keys on OTel v2, Langfuse logging the OTel v2 fan-out does not carry still reaches the +tenant's own Langfuse project: a key's failure-only entry on a failed call, a config ``default_team_settings`` entry, +and a key destination's call the proxy rejected before any provider. A ``success_and_failure`` key entry sends that +project exactly one copy of a failed call.""" + +from __future__ import annotations + +import json +import uuid +from collections.abc import Callable, Iterator +from typing import Final, Literal + +import pytest +from integration._support.client import Gateway, gateway_from_environment, held +from integration._support.otlp_sink import ( + AuditConfigWriter, + Span, + SpanSinks, + langfuse_callback_variables, + recorded_spans, + spans_for_trace, +) +from integration._support.process import owned_proxy_process +from integration._support.wire import Reply, Request, Wire, wire_server +from pydantic import JsonValue + +OPERATION: Final = "gen_ai.operation.name" +TAGS: Final = ["agent:orders", "run:r1"] +CONFIG_TEAM: Final = f"lf-config-team-{uuid.uuid4().hex[:8]}" +FAIL: Final = "please-fail" + + +def _chat(request: Request) -> Reply: + if FAIL.encode() in request.body: + return Reply( + status=500, body=json.dumps({"error": {"message": "upstream boom", "type": "server_error"}}).encode() + ) + return Reply( + body=json.dumps( + { + "id": f"chatcmpl-{uuid.uuid4().hex}", + "object": "chat.completion", + "created": 1, + "model": "gpt-4o-mini", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 7, "completion_tokens": 2, "total_tokens": 9}, + } + ).encode() + ) + + +@pytest.fixture(scope="module") +def provider() -> Iterator[Wire]: + with wire_server(_chat) as wire: + yield wire + + +@pytest.fixture(scope="module") +def proxy( + audit_sinks: SpanSinks, + otel_audit_config: AuditConfigWriter, + tmp_path_factory: pytest.TempPathFactory, +) -> Iterator[Gateway]: + directory: Final = tmp_path_factory.mktemp("langfuse-legacy-destinations") + settings: Final[dict[str, JsonValue]] = { + "callbacks": [], + "default_team_settings": [ + { + "team_id": CONFIG_TEAM, + "success_callback": ["langfuse_otel"], + "callback_vars": dict(langfuse_callback_variables(audit_sinks)), + } + ], + } + with ( + gateway_from_environment() as gateway, + owned_proxy_process( + gateway, + directory, + { + "LITELLM_OTEL_V2": "1", + "OTEL_BSP_SCHEDULE_DELAY": "300", + "OTEL_EXPORTER": "otlp_http", + "OTEL_ENDPOINT": audit_sinks.operator, + }, + config=otel_audit_config(directory, settings), + remove_environment=("LANGFUSE_PUBLIC_KEY", "LANGFUSE_SECRET_KEY", "LANGFUSE_HOST"), + workers=2, + ) as owned, + ): + yield owned.gateway + + +def _tenant_spans(url: str, since: int, trace_id: str) -> tuple[Span, ...]: + return spans_for_trace(recorded_spans(url, since)[1], trace_id) + + +def _tags(value: object) -> object: + """The tags a span carries: the legacy logger writes them as a JSON string, OTel v2 as a list.""" + return json.loads(value) if isinstance(value, str) else value + + +def _settled(spans: tuple[Span, ...]) -> bool: + """The model span and the request's root span have both arrived.""" + return any(OPERATION in span["attributes"] for span in spans) and any( + str(span["name"]).startswith("POST ") for span in spans + ) + + +def _logging_entry(audit_sinks: SpanSinks, callback_type: str) -> list[JsonValue]: + return [ + { + "callback_name": "langfuse_otel", + "callback_type": callback_type, + "callback_vars": langfuse_callback_variables(audit_sinks), + } + ] + + +def _chat_status(proxy: Gateway, model: str, key: str, trace_id: str, content: str) -> int: + body: Final[dict[str, JsonValue]] = { + "model": model, + "messages": [{"role": "user", "content": content}], + "metadata": {"tags": [*TAGS], "session_id": f"sess-{trace_id}", "trace_user_id": "user-42"}, + } + return proxy.request( + "POST", + "/v1/chat/completions", + body, + key=key, + headers={"traceparent": f"00-{trace_id}-{uuid.uuid4().hex[:16]}-01"}, + ).status_code + + +def _trace_fields(spans: tuple[Span, ...]) -> list[tuple[object, object]]: + return [ + (span["attributes"]["user.id"], _tags(span["attributes"].get("langfuse.trace.tags"))) + for span in spans + if "user.id" in span["attributes"] + ] + + +def _carries_the_trace_fields_once(spans: tuple[Span, ...]) -> bool: + return _trace_fields(spans) == [("user-42", TAGS)] + + +def _one_failed_call(spans: tuple[Span, ...]) -> bool: + return ( + _settled(spans) + and len([span for span in spans if OPERATION in span["attributes"]]) == 1 + and "litellm_request" not in {span["name"] for span in spans} + ) + + +def _send( + proxy: Gateway, + provider: Wire, + audit_sinks: SpanSinks, + setup: Literal["failure", "config_team", "both"], + expected: Callable[[tuple[Span, ...]], bool], +) -> tuple[int, tuple[Span, ...]]: + logging_entry: Final = _logging_entry(audit_sinks, "failure" if setup == "failure" else "success_and_failure") + with proxy.scenario() as scenario: + model: Final = scenario.model(api_base=provider.url + "/v1", num_retries=0) + alias: Final = f"lf-legacy-{setup}-{uuid.uuid4().hex}" + key: Final = ( + scenario.key(key_alias=alias, team_id=scenario.team(team_id=CONFIG_TEAM)) + if setup == "config_team" + else scenario.key(key_alias=alias, metadata={"logging": logging_entry}) + ) + since: Final = recorded_spans(audit_sinks.tenant)[0] + trace_id: Final = uuid.uuid4().hex + content: Final = f"lf-legacy-{setup}-{trace_id}" + ("" if setup == "config_team" else f" {FAIL}") + status: Final = _chat_status(proxy, model, key, trace_id, content) + tenant: Final = held( + lambda: _tenant_spans(audit_sinks.tenant, since, trace_id), expected, holding=3, seconds=40 + ) + return status, tenant + + +@pytest.mark.parametrize("setup", ["failure", "config_team"]) +def test_a_tenant_langfuse_project_without_a_destination_still_gets_the_trace( + setup: Literal["failure", "config_team"], proxy: Gateway, provider: Wire, audit_sinks: SpanSinks +) -> None: + status, tenant = _send(proxy, provider, audit_sinks, setup, _carries_the_trace_fields_once) + + assert status == (500 if setup == "failure" else 200) + assert _trace_fields(tenant) == [("user-42", TAGS)], [span["name"] for span in tenant] + + +def test_a_key_destinations_call_rejected_before_any_provider_still_gets_the_trace_fields( + proxy: Gateway, provider: Wire, audit_sinks: SpanSinks +) -> None: + with proxy.scenario() as scenario: + model: Final = scenario.model(api_base=provider.url + "/v1", num_retries=0) + key: Final = scenario.key( + key_alias=f"lf-legacy-rejected-{uuid.uuid4().hex}", + metadata={"logging": _logging_entry(audit_sinks, "success_and_failure")}, + rpm_limit=1, + ) + since: Final = recorded_spans(audit_sinks.tenant)[0] + statuses: Final = { + trace_id: _chat_status(proxy, model, key, trace_id, f"lf-legacy-rejected-{trace_id}") + for trace_id in [uuid.uuid4().hex for _ in range(4)] + } + rejected: Final = [trace_id for trace_id, status in statuses.items() if status == 429] + assert rejected, statuses + tenant: Final = held( + lambda: _tenant_spans(audit_sinks.tenant, since, rejected[0]), + _carries_the_trace_fields_once, + holding=3, + seconds=40, + ) + + assert _trace_fields(tenant) == [("user-42", TAGS)], [span["name"] for span in tenant] + + +def test_a_failed_call_reaches_a_success_and_failure_destination_once( + proxy: Gateway, provider: Wire, audit_sinks: SpanSinks +) -> None: + status, tenant = _send(proxy, provider, audit_sinks, "both", _one_failed_call) + + assert status == 500 + assert _one_failed_call(tenant), [span["name"] for span in tenant] diff --git a/tests/integration/observability/test_langfuse_otel_tenant_only.py b/tests/integration/observability/test_langfuse_otel_tenant_only.py new file mode 100644 index 00000000000..4204d286d82 --- /dev/null +++ b/tests/integration/observability/test_langfuse_otel_tenant_only.py @@ -0,0 +1,182 @@ +"""A key's or team's own Langfuse project gets one copy of the call with the Langfuse trace fields when the operator +has no Langfuse keys, whether or not ``langfuse_otel`` is listed in ``litellm_settings.callbacks``.""" + +from __future__ import annotations + +import json +import uuid +from collections.abc import Iterator +from typing import Final, Literal, cast + +import pytest +from integration._support.client import Gateway, gateway_from_environment, held +from integration._support.otlp_sink import ( + AuditConfigWriter, + Span, + SpanSinks, + langfuse_callback_variables, + recorded_spans, + spans_for_trace, +) +from integration._support.process import owned_proxy_process +from integration._support.wire import Reply, Request, Wire, wire_server +from pydantic import JsonValue + +OPERATION: Final = "gen_ai.operation.name" +TAGS: Final = ["agent:orders", "run:r1"] +LEGACY_SPANS: Final = frozenset({"litellm_request", "raw_gen_ai_request"}) + + +def _chat(_request: Request) -> Reply: + return Reply( + body=json.dumps( + { + "id": f"chatcmpl-{uuid.uuid4().hex}", + "object": "chat.completion", + "created": 1, + "model": "gpt-4o-mini", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 7, "completion_tokens": 2, "total_tokens": 9}, + } + ).encode() + ) + + +@pytest.fixture(scope="module") +def provider() -> Iterator[Wire]: + with wire_server(_chat) as wire: + yield wire + + +@pytest.fixture(scope="module", params=[["langfuse_otel"], []], ids=["listed", "not-listed"]) +def proxy( + request: pytest.FixtureRequest, + audit_sinks: SpanSinks, + otel_audit_config: AuditConfigWriter, + tmp_path_factory: pytest.TempPathFactory, +) -> Iterator[Gateway]: + directory: Final = tmp_path_factory.mktemp("langfuse-tenant-only") + with ( + gateway_from_environment() as gateway, + owned_proxy_process( + gateway, + directory, + { + "LITELLM_OTEL_V2": "1", + "OTEL_BSP_SCHEDULE_DELAY": "300", + "OTEL_EXPORTER": "otlp_http", + "OTEL_ENDPOINT": audit_sinks.operator, + }, + config=otel_audit_config(directory, {"callbacks": cast("list[JsonValue]", request.param)}), + remove_environment=("LANGFUSE_PUBLIC_KEY", "LANGFUSE_SECRET_KEY", "LANGFUSE_HOST"), + workers=2, + ) as owned, + ): + yield owned.gateway + + +def _trace_spans(url: str, since: int, trace_id: str) -> tuple[Span, ...]: + return spans_for_trace(recorded_spans(url, since)[1], trace_id) + + +def _model_spans(spans: tuple[Span, ...]) -> tuple[Span, ...]: + return tuple(span for span in spans if OPERATION in span["attributes"]) + + +def _one_model_span(spans: tuple[Span, ...]) -> bool: + return len(_model_spans(spans)) == 1 and not LEGACY_SPANS & {str(span["name"]) for span in spans} + + +@pytest.mark.parametrize("level", ["key", "team"]) +def test_a_tenant_langfuse_project_gets_user_session_and_tags_without_operator_keys( + level: Literal["key", "team"], proxy: Gateway, provider: Wire, audit_sinks: SpanSinks +) -> None: + logging: Final[list[JsonValue]] = [ + { + "callback_name": "langfuse_otel", + "callback_type": "success", + "callback_vars": langfuse_callback_variables(audit_sinks), + } + ] + with proxy.scenario() as scenario: + model: Final = scenario.model(api_base=provider.url + "/v1") + alias: Final = f"lf-tenant-only-{uuid.uuid4().hex}" + key: Final = ( + scenario.key(key_alias=alias, metadata={"logging": logging}) + if level == "key" + else scenario.key(key_alias=alias, team_id=scenario.team(metadata={"logging": logging})) + ) + since: Final = (recorded_spans(audit_sinks.tenant)[0], recorded_spans(audit_sinks.operator)[0]) + trace_id: Final = uuid.uuid4().hex + session: Final = f"sess-{trace_id}" + body: Final[dict[str, JsonValue]] = { + "model": model, + "messages": [{"role": "user", "content": f"lf-tenant-only-{trace_id}"}], + "metadata": {"tags": [*TAGS], "session_id": session, "trace_user_id": "user-42"}, + } + response: Final = proxy.request( + "POST", + "/v1/chat/completions", + body, + key=key, + headers={"traceparent": f"00-{trace_id}-{uuid.uuid4().hex[:16]}-01"}, + ) + assert response.status_code == 200, response.text + + tenant: Final = _model_spans( + held(lambda: _trace_spans(audit_sinks.tenant, since[0], trace_id), _one_model_span, holding=3, seconds=40) + ) + operator: Final = _model_spans( + held(lambda: _trace_spans(audit_sinks.operator, since[1], trace_id), _one_model_span, holding=3, seconds=40) + ) + + assert [ + {name: span["attributes"].get(name) for name in ("user.id", "session.id", "langfuse.trace.tags")} + for span in tenant + ] == [{"user.id": "user-42", "session.id": session, "langfuse.trace.tags": TAGS}], [s["name"] for s in tenant] + assert "langfuse.observation.metadata" in tenant[0]["attributes"], sorted(tenant[0]["attributes"]) + assert f"lf-tenant-only-{trace_id}" in str(tenant[0]["attributes"].get("langfuse.observation.input")), sorted( + tenant[0]["attributes"] + ) + assert len(operator) == 1, [span["name"] for span in operator] + assert f"lf-tenant-only-{trace_id}" not in json.dumps(operator[0]["attributes"]), sorted(operator[0]["attributes"]) + assert [ + name for name in operator[0]["attributes"] if name in ("user.id", "session.id") or name.startswith("langfuse.") + ] == [], sorted(operator[0]["attributes"]) + + +def test_a_tenant_langfuse_project_gets_the_keys_langfuse_environment( + proxy: Gateway, provider: Wire, audit_sinks: SpanSinks +) -> None: + logging: Final[list[JsonValue]] = [ + { + "callback_name": "langfuse_otel", + "callback_type": "success", + "callback_vars": {**langfuse_callback_variables(audit_sinks), "langfuse_environment": "prod"}, + } + ] + with proxy.scenario() as scenario: + model: Final = scenario.model(api_base=provider.url + "/v1") + key: Final = scenario.key(key_alias=f"lf-tenant-env-{uuid.uuid4().hex}", metadata={"logging": logging}) + since: Final = (recorded_spans(audit_sinks.tenant)[0], recorded_spans(audit_sinks.operator)[0]) + trace_id: Final = uuid.uuid4().hex + response: Final = proxy.request( + "POST", + "/v1/chat/completions", + {"model": model, "messages": [{"role": "user", "content": f"lf-tenant-env-{trace_id}"}]}, + key=key, + headers={"traceparent": f"00-{trace_id}-{uuid.uuid4().hex[:16]}-01"}, + ) + assert response.status_code == 200, response.text + + tenant: Final = held( + lambda: _trace_spans(audit_sinks.tenant, since[0], trace_id), _one_model_span, holding=3, seconds=40 + ) + operator: Final = held( + lambda: _trace_spans(audit_sinks.operator, since[1], trace_id), _one_model_span, holding=3, seconds=40 + ) + + assert [span["resource"].get("langfuse.environment") for span in tenant] == ["prod"] * len(tenant), [ + span["name"] for span in tenant + ] + assert not any("langfuse.environment" in span["resource"] for span in operator), [span["name"] for span in operator] diff --git a/tests/unit/integrations/otel/test_otel_v2_destinations.py b/tests/unit/integrations/otel/test_otel_v2_destinations.py index 78dcb7bc8f1..eb422508789 100644 --- a/tests/unit/integrations/otel/test_otel_v2_destinations.py +++ b/tests/unit/integrations/otel/test_otel_v2_destinations.py @@ -17,7 +17,7 @@ import pytest from opentelemetry import trace as trace_api from opentelemetry.context import Context from opentelemetry.sdk.resources import Resource -from opentelemetry.sdk.trace import ReadableSpan, TracerProvider +from opentelemetry.sdk.trace import ReadableSpan, SpanLimits, TracerProvider from opentelemetry.sdk.trace.export import SimpleSpanProcessor from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter from opentelemetry.sdk.trace.id_generator import RandomIdGenerator @@ -27,6 +27,7 @@ import litellm from litellm.integrations.custom_logger import CustomLogger from litellm.integrations.otel import GenAIOperation from litellm.integrations.otel import logger as otel_logger +from litellm.integrations.otel.emitter import SpanEmitter from litellm.integrations.otel.logger import ( OpenTelemetryV2, build_otel_v2_logger, @@ -34,6 +35,7 @@ from litellm.integrations.otel.logger import ( publish_global_otel_v2_provider, ) from litellm.integrations.otel.mappers import resolve_mappers +from litellm.integrations.otel.mappers.langfuse import LANGFUSE_REQUEST_METADATA_ATTRIBUTES, LangfuseMapper from litellm.integrations.otel.model.config import ( CaptureMessageContent, ExporterOwner, @@ -52,6 +54,8 @@ from litellm.integrations.otel.model.payloads import ( ServerInfo, ToolDefinition, ) +from litellm.integrations.otel.model.spans import SpanRole +from litellm.integrations.otel.model.trace_controls import TraceControls from litellm.integrations.otel.plumbing import providers as otel_providers from litellm.integrations.otel.plumbing.context import ( destination_backends, @@ -3435,6 +3439,276 @@ class TestArizeTenantOnly: assert loggers == [at_startup] +class TestLangfuseDestinationFields: + """A key's or team's Langfuse destination gets the Langfuse fields from whichever logger the proxy published.""" + + def test_a_key_destination_gets_the_langfuse_fields_from_the_generic_startup_logger( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + """With ``langfuse_otel`` only in a key's or team's logging, the startup logger is the + generic one; the key's project still gets the user, session and tags, and the + operator's stdout placeholder still gets no Langfuse request metadata.""" + credential_less_proxy(monkeypatch) + capture: Final = _ExporterCapture() + for kind in ("otlp_grpc", "otlp_http", "console"): + monkeypatch.setitem(otel_providers._EXPORTER_FACTORIES, kind, capture.build) # pyright: ignore[reportPrivateUsage] # the factory table is the seam for capturing each exporter + config: Final = OpenTelemetryV2Config() + provider: Final = build_tracer_provider(config, tenant_overrides=True) + attach_tenant_fan_out(provider, config) + emitter: Final = SpanEmitter(get_tracer(provider, "litellm"), config) + call: Final = replace( + _MODEL_CALL_WITHOUT_CONTENT, + trace=TraceControls(user_id="user-42", session_id="sess-1", tags=("agent:orders",)), + request_metadata=MappingProxyType({"user_api_key_alias": "prod-key"}), + ) + + def with_destination() -> None: + set_request_destinations(deliverable_destinations((LANGFUSE_DEST,), provider)) + emitter.emit(SpanRole.LLM_CALL, call) + + in_fresh_context(with_destination) + provider.force_flush() + + received: Final = { + headers: exporter.get_finished_spans()[0].attributes or {} for headers, exporter in capture.built + } + to_langfuse: Final = received[LANGFUSE_DEST.header_string()] + assert {key: to_langfuse.get(key) for key in ("user.id", "session.id", "langfuse.trace.tags")} == { + "user.id": "user-42", + "session.id": "sess-1", + "langfuse.trace.tags": ("agent:orders",), + } + assert "langfuse.observation.metadata" in to_langfuse + assert "langfuse.observation.metadata" not in received[None] + + @staticmethod + def _received_by_account( + monkeypatch: pytest.MonkeyPatch, + config: OpenTelemetryV2Config, + destinations: tuple[OtelDestination, ...], + with_content: bool = False, + ) -> Mapping[str | None, Mapping[str, object]]: + """The attributes of the model call each account received, keyed by the headers it was sent with.""" + credential_less_proxy(monkeypatch) + capture: Final = _ExporterCapture() + for kind in ("otlp_grpc", "otlp_http", "console"): + monkeypatch.setitem(otel_providers._EXPORTER_FACTORIES, kind, capture.build) # pyright: ignore[reportPrivateUsage] # the factory table is the seam for capturing each exporter + provider: Final = build_tracer_provider(config, tenant_overrides=True) + attach_tenant_fan_out(provider, config) + emitter: Final = SpanEmitter(get_tracer(provider, "litellm"), config) + call: Final = replace( + _MODEL_CALL_WITH_CONTENT if with_content else _MODEL_CALL_WITHOUT_CONTENT, + trace=TraceControls(user_id="user-42", session_id="sess-1", tags=("agent:orders",)), + request_metadata=MappingProxyType({"user_api_key_alias": "prod-key"}), + ) + + def with_destinations() -> None: + set_request_destinations(deliverable_destinations(destinations, provider)) + emitter.emit(SpanRole.LLM_CALL, call) + + in_fresh_context(with_destinations) + provider.force_flush() + return MappingProxyType( + {headers: dict(exporter.get_finished_spans()[0].attributes or {}) for headers, exporter in capture.built} + ) + + def test_only_the_langfuse_destination_gets_the_langfuse_fields_written_for_it( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + """The operator's own exporter and the key's Arize destination get the model call as they would with no + Langfuse destination on the request.""" + received: Final = self._received_by_account(monkeypatch, OpenTelemetryV2Config(), (LANGFUSE_DEST, ARIZE_DEST)) + + assert received[LANGFUSE_DEST.header_string()]["user.id"] == "user-42" + assert { + headers: sorted(LangfuseMapper.LLM_CALL_ATTRIBUTES & attributes.keys()) + for headers, attributes in received.items() + if headers != LANGFUSE_DEST.header_string() + } == {None: [], ARIZE_DEST.header_string(): []} + + def test_an_exporter_whose_config_maps_langfuse_keeps_the_langfuse_fields_but_the_request_metadata( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + received: Final = self._received_by_account( + monkeypatch, OpenTelemetryV2Config(mapper_names=["genai", "langfuse"]), (LANGFUSE_DEST,) + ) + + assert {key: received[None].get(key) for key in ("user.id", "session.id", "langfuse.trace.tags")} == { + "user.id": "user-42", + "session.id": "sess-1", + "langfuse.trace.tags": ("agent:orders",), + } + assert not LANGFUSE_REQUEST_METADATA_ATTRIBUTES & received[None].keys() + + @pytest.mark.parametrize( + ("key_environment", "tracing_environment", "expected"), + [ + ("prod", None, {"langfuse.environment": "prod"}), + ("prod", "staging", {"langfuse.environment": "prod"}), + (None, "staging", {"langfuse.environment": "staging"}), + (None, None, {}), + ], + ids=["key", "key-over-env", "env", "neither"], + ) + def test_a_langfuse_destination_names_the_environment_the_legacy_logger_wrote( + self, + monkeypatch: pytest.MonkeyPatch, + key_environment: str | None, + tracing_environment: str | None, + expected: Mapping[str, str], + ) -> None: + credential_less_proxy(monkeypatch) + if tracing_environment is None: + monkeypatch.delenv("LANGFUSE_TRACING_ENVIRONMENT", raising=False) + else: + monkeypatch.setenv("LANGFUSE_TRACING_ENVIRONMENT", tracing_environment) + + destination: Final = destination_for( + "langfuse_otel", + StandardCallbackDynamicParams( + langfuse_public_key="pk", langfuse_secret_key="sk", langfuse_environment=key_environment + ), + ) + + assert destination is not None + assert dict(destination.resource_attributes) == expected + + def test_a_langfuse_destination_on_a_proxy_with_langfuse_keys_names_no_environment( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + """With operator keys the older Langfuse logger never runs, so there is no environment of its to keep.""" + monkeypatch.setenv("LANGFUSE_PUBLIC_KEY", "pk-operator") + monkeypatch.setenv("LANGFUSE_SECRET_KEY", "sk-operator") + + destination: Final = destination_for( + "langfuse_otel", + StandardCallbackDynamicParams( + langfuse_public_key="pk", langfuse_secret_key="sk", langfuse_environment="prod" + ), + ) + + assert destination is not None + assert dict(destination.resource_attributes) == {} + + @pytest.mark.parametrize( + ("configured", "operator_keys", "expected"), + [ + (None, False, "span_only"), + ("no_content", False, "no_content"), + (None, True, None), + ("span_only", True, "span_only"), + ], + ids=["unset", "key-no-content", "operator-keys", "operator-keys-key-span-only"], + ) + def test_a_langfuse_destination_captures_the_prompt_where_the_legacy_logger_sent_it( + self, + monkeypatch: pytest.MonkeyPatch, + configured: str | None, + operator_keys: bool, + expected: str | None, + ) -> None: + credential_less_proxy(monkeypatch) + if operator_keys: + monkeypatch.setenv("LANGFUSE_PUBLIC_KEY", "pk-operator") + monkeypatch.setenv("LANGFUSE_SECRET_KEY", "sk-operator") + + destination: Final = destination_for( + "langfuse_otel", + StandardCallbackDynamicParams(langfuse_public_key="pk", langfuse_secret_key="sk"), + capture_message_content=configured, + ) + + assert destination is not None + assert destination.capture_message_content == expected + + def test_only_the_langfuse_destination_gets_the_prompt_it_captures_by_default( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + credential_less_proxy(monkeypatch) + destination: Final = destination_for( + "langfuse_otel", StandardCallbackDynamicParams(langfuse_public_key="pk", langfuse_secret_key="sk") + ) + arize: Final = destination_for("arize", StandardCallbackDynamicParams(arize_space_id="s", arize_api_key="k")) + assert destination is not None + assert arize is not None + + received: Final = self._received_by_account( + monkeypatch, OpenTelemetryV2Config(), (destination, arize), with_content=True + ) + + assert {headers: "langfuse.observation.input" in attributes for headers, attributes in received.items()} == { + destination.header_string(): True, + arize.header_string(): False, + None: False, + } + assert not [key for key in received[None] if key.startswith("gen_ai.input")] + + def test_only_the_langfuse_destination_gets_the_keys_environment(self, monkeypatch: pytest.MonkeyPatch) -> None: + credential_less_proxy(monkeypatch) + capture: Final = _ExporterCapture() + for kind in ("otlp_grpc", "otlp_http", "console"): + monkeypatch.setitem(otel_providers._EXPORTER_FACTORIES, kind, capture.build) # pyright: ignore[reportPrivateUsage] # the factory table is the seam for capturing each exporter + config: Final = OpenTelemetryV2Config() + provider: Final = build_tracer_provider(config, tenant_overrides=True) + attach_tenant_fan_out(provider, config) + emitter: Final = SpanEmitter(get_tracer(provider, "litellm"), config) + destination: Final = destination_for( + "langfuse_otel", {"langfuse_public_key": "pk", "langfuse_secret_key": "sk", "langfuse_environment": "prod"} + ) + assert destination is not None + + def with_destination() -> None: + set_request_destinations(deliverable_destinations((destination,), provider)) + emitter.emit(SpanRole.LLM_CALL, _MODEL_CALL_WITHOUT_CONTENT) + + in_fresh_context(with_destination) + provider.force_flush() + + environments: Final = { + headers: exporter.get_finished_spans()[0].resource.attributes.get("langfuse.environment") + for headers, exporter in capture.built + } + assert environments == {destination.header_string(): "prod", None: None} + + @staticmethod + def _attributes( + destinations: tuple[OtelDestination, ...], limit: int | None, call: LLMCallSpanData + ) -> dict[str, object]: + exporter: Final = InMemorySpanExporter() + provider: Final = TracerProvider(span_limits=SpanLimits(max_span_attributes=limit)) + provider.add_span_processor(SimpleSpanProcessor(exporter)) + emitter: Final = SpanEmitter(get_tracer(provider, "litellm"), OpenTelemetryV2Config()) + + def run() -> None: + set_request_destinations(destinations) + emitter.emit(SpanRole.LLM_CALL, call) + + in_fresh_context(run) + return dict(exporter.get_finished_spans()[0].attributes or {}) + + def test_a_span_at_its_attribute_limit_sheds_the_destinations_langfuse_fields_first(self) -> None: + generic: Final = self._attributes((), None, _MODEL_CALL_WITH_CONTENT) + + assert "user.id" not in generic + assert self._attributes((LANGFUSE_DEST,), len(generic), _MODEL_CALL_WITH_CONTENT) == generic, ( + "the generic attributes all survive" + ) + + def test_a_span_at_its_attribute_limit_sheds_the_destinations_request_metadata_before_its_other_fields( + self, + ) -> None: + call: Final = replace( + _MODEL_CALL_WITH_CONTENT, request_metadata=MappingProxyType({"user_api_key_alias": "prod-key"}) + ) + whole: Final = self._attributes((LANGFUSE_DEST,), None, call) + + squeezed: Final = self._attributes((LANGFUSE_DEST,), len(whole) - 1, call) + + assert len(squeezed) == len(whole) - 1 + assert whole.keys() - squeezed.keys() <= LANGFUSE_REQUEST_METADATA_ATTRIBUTES + assert squeezed["langfuse.observation.type"] == "generation" + + class TestContextIsolation: def test_destinations_do_not_leak_between_requests(self): def first(): @@ -5103,7 +5377,10 @@ class TestCaptureMessageContent: def test_the_setting_rides_the_teams_destination_and_omission_stays_omitted( self, monkeypatch: pytest.MonkeyPatch, setting: str | None ) -> None: + """On a proxy with Langfuse keys of its own; without them see ``TestLangfuseDestinationFields``.""" monkeypatch.setenv("LITELLM_OTEL_V2", "true") + monkeypatch.setenv("LANGFUSE_PUBLIC_KEY", "pk-operator") + monkeypatch.setenv("LANGFUSE_SECRET_KEY", "sk-operator") is_otel_v2_enabled.cache_clear() callback_vars: Final = { "langfuse_public_key": "pk-team", diff --git a/tests/unit/integrations/test_langfuse_otel.py b/tests/unit/integrations/test_langfuse_otel.py index 5d37a145de6..a2803719c75 100644 --- a/tests/unit/integrations/test_langfuse_otel.py +++ b/tests/unit/integrations/test_langfuse_otel.py @@ -1,9 +1,12 @@ +import asyncio import base64 +import contextvars import json import os from collections.abc import Mapping, Sequence from dataclasses import dataclass from datetime import datetime, timezone +from types import MappingProxyType from typing import Final from unittest.mock import MagicMock, patch @@ -12,12 +15,17 @@ from openai.types.responses import ResponseFunctionToolCall from opentelemetry import trace from opentelemetry.context import Context from opentelemetry.sdk.trace import SpanLimits, TracerProvider +from opentelemetry.sdk.trace.export import SimpleSpanProcessor +from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter from opentelemetry.trace import Span import litellm -from litellm.constants import REDACTED_BY_LITELLM +from litellm.constants import LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL, REDACTED_BY_LITELLM from litellm.integrations.langfuse.langfuse_otel import LangfuseOtelLogger from litellm.integrations.opentelemetry import LITELLM_PROXY_REQUEST_SPAN_NAME, OpenTelemetry, OpenTelemetryConfig +from litellm.integrations.otel.model.destination import OtelDestination +from litellm.integrations.otel.plumbing.context import set_request_destinations +from litellm.proxy._types import UserAPIKeyAuth from litellm.types.llms.openai import ResponsesAPIResponse from litellm.types.utils import Choices, ModelResponse, StandardCallbackDynamicParams @@ -1507,3 +1515,75 @@ def test_langfuse_otel_does_not_start_proxy_request_span() -> None: finally: langfuse_provider.shutdown() generic_provider.shutdown() + + +_HOOK_CALL: Final[Mapping[str, object]] = MappingProxyType({}) + + +def _spans_the_legacy_logger_exports( + destinations: tuple[OtelDestination, ...], call: Mapping[str, object] = _HOOK_CALL +) -> tuple[str, ...]: + exporter: Final = InMemorySpanExporter() + provider: Final = TracerProvider() + provider.add_span_processor(SimpleSpanProcessor(exporter)) + with patch.dict(os.environ, LANGFUSE_ENV_ONLY, clear=True): + logger: Final = LangfuseOtelLogger(tracer_provider=provider) + at: Final = datetime(2026, 1, 1, tzinfo=timezone.utc) + kwargs: Final[dict[str, object]] = {**CHAT_KWARGS, **call, "litellm_params": {"metadata": {}}} + server_span: Final = provider.get_tracer("proxy").start_span("POST /v1/chat/completions") + + async def log() -> None: + set_request_destinations(destinations) + logger.log_success_event(dict(kwargs), _chat_response("ok"), at, at) # pyright: ignore[reportUnknownMemberType] # the v1 logger hooks are untyped + await logger.async_log_failure_event({**kwargs, "exception": ValueError("boom")}, None, at, at) # pyright: ignore[reportUnknownMemberType] # the v1 logger hooks are untyped + await logger.async_post_call_failure_hook({}, ValueError("boom"), UserAPIKeyAuth(parent_otel_span=server_span)) + + try: + contextvars.copy_context().run(asyncio.run, log()) + return tuple(span.name for span in exporter.get_finished_spans()) + finally: + provider.shutdown() + + +_LANGFUSE_DESTINATION: Final = OtelDestination( + endpoint="http://tenant.local/api/public/otel", + headers={"Authorization": "Basic dGVuYW50"}, + callback_name="langfuse_otel", +) + + +_PROVIDER_CALL: Final[Mapping[str, object]] = MappingProxyType( + {"standard_logging_object": {"id": "c1"}, "api_call_start_time": datetime(2026, 1, 1, tzinfo=timezone.utc)} +) + + +@pytest.mark.parametrize("call", [_HOOK_CALL, _PROVIDER_CALL], ids=["hook", "provider-call"]) +def test_a_request_with_a_langfuse_destination_exports_nothing_through_the_legacy_logger( + call: Mapping[str, object], +) -> None: + assert _spans_the_legacy_logger_exports((_LANGFUSE_DESTINATION,), call) == () + + +def test_a_call_rejected_before_any_provider_still_exports_through_the_legacy_logger() -> None: + rejected: Final = {"standard_logging_object": {"id": "c1"}, LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL: True} + + assert _spans_the_legacy_logger_exports((_LANGFUSE_DESTINATION,), rejected) == ( + "litellm_request", + "raw_gen_ai_request", + "litellm_request", + ) + + +def test_a_request_without_a_langfuse_destination_still_exports_through_the_legacy_logger() -> None: + logged: Final = ( + "litellm_request", + "raw_gen_ai_request", + "litellm_request", + "Failed Proxy Server Request", + "POST /v1/chat/completions", + ) + + assert _spans_the_legacy_logger_exports(()) == logged + assert _spans_the_legacy_logger_exports((_LANGFUSE_DESTINATION.model_copy(update={"callback_name": "arize"}),)) == ( + logged + )